FEATURE EXTRACTION IN NDE WELD IMAGES

Abstract

Welding is the most versatile way of permanently joining two metal plates. In this process, heat is applied to the metal pieces to be joined; the metal pieces melt in resultant and fuse together to form a permanent bond. Welding plays a major role in industries for the purpose of construction, joining and repairing of steel beams, reinforcing rods in buildings, bridges, spacecraft, pipe lines, nuclear containers etc. Welding, in its various processes and methods, can produce a number of different types of discontinuities, which can result, from material inconsistencies, operator error, or from uncontrollable factors. Regardless of the source of error, detection of discontinuities is critical. An unsatisfactory weld will seriously reduce the bond between two materials and may cause failure. These flaws are seen in the form of gaps, crack or bubbles. Common weld defects include lack of fusion, lack of penetration or excess penetration, gas cavities, slag inclusions, cracks, undercuts, lamellar tearing, shrinking cavities etc. To identify these flaws, Non Destructive Inspection (NDI) or Testing (NDT) is used. The NDI is the examination of an object or material with technology that retains its future usefulness without destroying or damaging the product or material. Commonly used non-destructive inspection methods include liquid penetrant, magnetic particle, eddy current, radiographic, and ultrasonic inspection. Welded steel plates are although, inspected by using many of the above NDT techniques, but X-ray inspection is used most often. UT is also used for imaging of flaws. The analysis of UT images may also be helpful in categorization of weld defects. The manual interpretation of UT and RT images depends upon the level of xxv expertise of the specialist and is usually a time consuming process as well as subjective. To remove the subjectivity in evaluation and expedite the inspection process, there is always a need for automation. Most of the NDT techniques have been made automated these days but the interpretation of signal still is not fully standardized. Image processing plays an important role to interpret these images with an aim of detecting flaws in weldments. With the help of image processing a semi-automated system can be developed, which would be more reliable as compared to the classical methods and also helpful in taking decision with the help of a set of manipulating tools. The present work has been an effort in the direction of automating the inspection process. The present research work is focused on detecting the weld defects in NDE weld images (X-ray and ultrasound). Specific aim here, has been to extract defect features once they have been detected. Feature extraction in present work has been performed in spatial domain. Various feature extraction algorithms have been used to determine features such as shape, size, and orientation of weld defects. To achieve the above objective, the weld image is segmented in different flaw regions. Once the flaw regions are identified, these are analyzed further to get the above mentioned features. This entire work has been broken in to three steps. The image enhancement forms the first step, which includes gray level and contrast manipulations, noise reduction, edge crispening and sharpening (or edge detection), interpolation and magnification, and pseudo coloring. The histogram equalization method has been applied here for improving the contrast of the image. This has been followed by noise removal. For noise removal, median filters and adaptive filters were used, which are very effective in reducing the impulsive noise and enhancing the contrast of the image. xxvi The segmentation ofthe images forms the second stage. In present work three segmentation techniques have been implemented. The implemented segmentation techniques are: edge-based, region growing and watershed transform. For implementing edge-based segmentation, edge detection algorithms such as, Sobel, Kirsch, Zero crossing and Canny are applied to determine the edges in the NDE X-ray images ofweldments. To select a suitable edge detection technique, these edge detection techniques were applied on variety of weld images both UT and RT. A comparative study ofperformance of these techniques on these images was performed which emphasizes that Canny operator and second order derivatives produce better results on most of these images. Based on the strength of edges in a specified local neighborhood, the confidence of each edge is either increased or decreased. Ifborder of region is not identifiable the border is traced using different morphological functions like filling holes, clear border, dilation, erosion, closing and opening with edge detectors and border-tracing technique. By all these techniques a close contour can be found which further gives an idea about the shape, size and area ofa particular flaw. The algorithms have been applied on the radiographic images of welds in steel collected from EURECTEST, International scientific Association Brussels, Belgium. Although, after morphological operations closed contours are obtained from the edgebased image segmentation technique, but in few cases closed contours are not formed. But the flaws present in the segmented images have been categorized successfully. The edge based segmentation technique works successfully on few types of flaws like slag inclusion, incomplete penetration and transverse cracks. The other segmentation techniques implemented for segmenting the images is region growing. For implementing, region-growing approach to segment X-ray xxvu weldment images, the selection of seed value is very important aspect to start the process. The seed/seeds value is obtained with the help ofhistogram plot ofthe image showing major valleys. The seed pixel grows into the region, based on the pre-defined homogeneity criteria ofgray level range in present case. This results in tracing ofclose boundary of the flaw. The region growing approach provides best results for lack of root penetration, undercuts and gas cavities. The third segmentation technique implemented is watershed transformation. The watershed algorithm is based on visualizing images in two spatial coordinates versus gray levels. To interpret topographically, three types of points are considered : (a) points belonging to a regional minimum; (b) points at which a drop of water, if placed at the location of any of those points, would fall with certainty to a single minimum; and (c) points at which water would be equally likely to fall to more than one such minimum for a particular regional minimum. The set of points satisfying the above point (b) are called catchment basin or watershed of that minimum and the points satisfying condition (c) form crest level on the topographic surface are termed as water shed lines or divide lines. To overcome the problem generated by watershed segmentation like poor detection of thin structures, poor detection of significant areas with low contrast boundaries, sensitivity to noise and over-segmentation, multistage watershed transformation has been developed and implemented inthis work. In this the region based post - preprocessing is applied with watershed segmentation for final contour identification. The watershed transformation is used twice to reduce noise , over segmentation and to obtain low contrast and thin lines. Watershed segmentation technique is best suited for identifying flaws of wormhole type gas cavities, lack of fusion, slag inclusion and slag line. XXVlll Defect feature extraction forms the third and final stage. Once the flaw regions are obtained after image segmentation, these flaw regions are analyzed for the determination of their features. The region labeling or identification is the first step. After labeling, the features such as shape, size defined in term of area, maximum length, maximum width and orientation are determined by using suitable algorithms. These parameters have been obtained in actual terms as the density of the image has been known here. The obtained results were compared with the results obtained from visual inspection. An experimental study has been performed at NDT Department of Bharat Heavy Electrical Ltd. (BHEL) at Hardwar for verification and suitability of the developed algorithms. BHEL Hardwar manufactures turbines, transformers and heavy electrical equipments. The NDT department is involved in the detection of flaws in weldments and categorizing them in to different groups. Categorization is aimed for quality control by grouping flaws into different categories on the basis of their approximate size. In this experimental study, different types of flaws were developed artificially in 20mm and 32mm thick mild steel surface using SMAW(Shielded Metal Arc Welding) and GMAW (Gas Metal Arc Welding) processes. The types of flaws, which were developed successfully, are lack of root penetration, slag inclusion, gas cavities/ porosity, and cracks. The approximate size and location of these flaws were physically determined by using air arc gouging and then implementing dye penetration. Radiographic and Ultrasonic NDT was performed on these artificially created flaws, which resulted in 2-D images and echo pattern of the weldments. The shape and size of the artificially developed flaws in the weldments were determined with the help of the above mentioned developed image processing algorithms on X-ray images. UT echo xxix patterns have been used for determining the depth of the flaws. A comparison between physically measured flaw information and that obtained from image processing algorithm was also performed. The results produced by the image processing algorithms have good correlation with the physically measured parameters. XXX